Decomposing 3D Shapes into Parts - Rutgers...
Transcript of Decomposing 3D Shapes into Parts - Rutgers...
Decomposing 3D Shapes intoParts
Christopher ToshPerceptual Science and Technology
Image Credit: Rong Liu
Motivation: Shape Analysis
• Applications in computergraphics, vision, andmany other fields needways of representingsurfaces in 3D.
Image credit: Attene et al.
Motivation: Shape Analysis
• Objects can be analyzedbetter when broken intopieces
• Studies suggest that wesee objects in terms oftheir parts
Image credit: Attene et al.
Starting Point: Segmentation• Segmentation is a partition of a surface.
• Ideally, those dividing lines are close to what humansnaturally perceive.
Image Credit: Attene et al.
Previous Work
3D Segmentation:• Shalfman et al. (2002)• Clustering
• Golovinskiy andFunkhouser (2008)• Graph Cuts
Image Credit: Papers cited
Segmentation Techniques• One approach is to map
surface points to some n-dimensional feature space.
• Then use a clustering algorithmto group the points. (k-means,mean shift, etc.)
• Problem: Hard to make bothfast and accurate.
Image Credit: Comaniciu and Meer TPAMI 2002
Segmentation Techniques
• More common approach is tobuild a connected graph whosevertices are points on thesurface.
• Edge weights are carefullychosen.
• Graph is broken into connectedcomponents in an optimal way.
Source: http://www-sipl.technion.ac.il/Info/Teaching_Projects_Text-Synt_e.shtml
From Segmentation to Parts
• Parts are the defining regions of a shape• Typically, elliptical• Held together by “glue”, which are the (usually)
hyperbolic connecting regions• Differ from segments in that not all of an object
must be considered to be in one part or another
Image Credit: Mi and DeCarlo ICCV 2007
2D Case: Already done
Mi and DeCarlo (2007)
• Used symmetry to locatepotential parts
• Separated 2D object intonatural parts.
Images Credit: Siddiqi and Kimia 1995; Mi and DeCarlo ICCV 2007
The Goal
• To design and implement aworkable algorithm that cansuccessfully decompose a3D object into parts
• To have those parts beapproximate to thoseperceived naturally byhumans
The Plan
• Build on existing segmentation algorithms toapproximate locations of parts
• Identify hyperbolic regions between segments• Fill in holes left behind by removing parts
Special thanks to:
• Matthew Stone
• Doug DeCarlo
• Kevin Sanik